4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #ifndef __LNET_TYPES_H__
34 #define __LNET_TYPES_H__
39 #include <linux/types.h>
40 /** \addtogroup lnet_addr
43 #define LNET_VERSION "0.7.0"
45 /** Portal reserved for LNet's own use.
46 * \see lustre/include/lustre/lustre_idl.h for Lustre portal assignments.
48 #define LNET_RESERVED_PORTAL 0
51 * Address of an end-point in an LNet network.
53 * A node can have multiple end-points and hence multiple addresses.
54 * An LNet network can be a simple network (e.g. tcp0) or a network of
55 * LNet networks connected by LNet routers. Therefore an end-point address
56 * has two parts: network ID, and address within a network.
58 * \see LNET_NIDNET, LNET_NIDADDR, and LNET_MKNID.
60 typedef __u64 lnet_nid_t;
63 * ID of a process in a node. Shortened as PID to distinguish from
64 * lnet_process_id, the global process ID.
66 typedef __u32 lnet_pid_t;
68 /** wildcard NID that matches any end-point address */
69 #define LNET_NID_ANY ((lnet_nid_t) -1)
70 /** wildcard PID that matches any lnet_pid_t */
71 #define LNET_PID_ANY ((lnet_pid_t) -1)
73 #define LNET_PID_RESERVED 0xf0000000 /* reserved bits in PID */
74 #define LNET_PID_USERFLAG 0x80000000 /* set in userspace peers */
75 #define LNET_PID_LUSTRE 12345
77 /* how an LNET NID encodes net:address */
78 /** extract the address part of an lnet_nid_t */
80 static inline __u32 LNET_NIDADDR(lnet_nid_t nid)
82 return nid & 0xffffffff;
85 static inline __u32 LNET_NIDNET(lnet_nid_t nid)
87 return (nid >> 32) & 0xffffffff;
90 static inline lnet_nid_t LNET_MKNID(__u32 net, __u32 addr)
92 return (((__u64)net) << 32) | addr;
95 static inline __u32 LNET_NETNUM(__u32 net)
100 static inline __u32 LNET_NETTYP(__u32 net)
102 return (net >> 16) & 0xffff;
105 static inline __u32 LNET_MKNET(__u32 type, __u32 num)
107 return (type << 16) | num;
110 #define WIRE_ATTR __attribute__((packed))
112 /* Packed version of struct lnet_process_id to transfer via network */
113 typedef struct lnet_process_id_packed {
115 lnet_pid_t pid; /* node id / process id */
116 } WIRE_ATTR lnet_process_id_packed;
118 /* The wire handle's interface cookie only matches one network interface in
119 * one epoch (i.e. new cookie when the interface restarts or the node
120 * reboots). The object cookie only matches one object on that interface
121 * during that object's lifetime (i.e. no cookie re-use). */
122 typedef struct lnet_handle_wire {
123 __u64 wh_interface_cookie;
124 __u64 wh_object_cookie;
125 } WIRE_ATTR lnet_handle_wire_t;
127 typedef enum lnet_msg_type {
135 /* The variant fields of the portals message header are aligned on an 8
136 * byte boundary in the message header. Note that all types used in these
137 * wire structs MUST be fixed size and the smaller types are placed at the
140 struct lnet_handle_wire dst_wmd;
146 struct lnet_handle_wire ack_wmd;
154 struct lnet_handle_wire return_wmd;
162 struct lnet_handle_wire dst_wmd;
170 typedef struct lnet_hdr {
175 __u32 type; /* enum lnet_msg_type */
176 __u32 payload_length; /* payload data to follow */
177 /*<------__u64 aligned------->*/
182 struct lnet_reply reply;
183 struct lnet_hello hello;
185 } WIRE_ATTR lnet_hdr_t;
187 /* A HELLO message contains a magic number and protocol version
188 * code in the header's dest_nid, the peer's NID in the src_nid, and
189 * LNET_MSG_HELLO in the type field. All other common fields are zero
190 * (including payload_size; i.e. no payload).
191 * This is for use by byte-stream LNDs (e.g. TCP/IP) to check the peer is
192 * running the same protocol and to find out its NID. These LNDs should
193 * exchange HELLO messages when a connection is first established. Individual
194 * LNDs can put whatever else they fancy in lnet_hdr::msg.
196 typedef struct lnet_magicversion {
197 __u32 magic; /* LNET_PROTO_TCP_MAGIC */
198 __u16 version_major; /* increment on incompatible change */
199 __u16 version_minor; /* increment on compatible change */
200 } WIRE_ATTR lnet_magic_version_t;
202 /* PROTO MAGIC for LNDs */
203 #define LNET_PROTO_IB_MAGIC 0x0be91b91
204 #define LNET_PROTO_GNI_MAGIC 0xb00fbabe /* ask Kim */
205 #define LNET_PROTO_TCP_MAGIC 0xeebc0ded
206 #define LNET_PROTO_ACCEPTOR_MAGIC 0xacce7100
207 #define LNET_PROTO_PING_MAGIC 0x70696E67 /* 'ping' */
209 /* Placeholder for a future "unified" protocol across all LNDs */
210 /* Current LNDs that receive a request with this magic will respond
211 * with a "stub" reply using their current protocol */
212 #define LNET_PROTO_MAGIC 0x45726963 /* ! */
214 #define LNET_PROTO_TCP_VERSION_MAJOR 1
215 #define LNET_PROTO_TCP_VERSION_MINOR 0
217 /* Acceptor connection request */
218 typedef struct lnet_acceptor_connreq {
219 __u32 acr_magic; /* PTL_ACCEPTOR_PROTO_MAGIC */
220 __u32 acr_version; /* protocol version */
221 __u64 acr_nid; /* target NID */
222 } WIRE_ATTR lnet_acceptor_connreq_t;
224 #define LNET_PROTO_ACCEPTOR_VERSION 1
226 typedef struct lnet_counters {
238 } WIRE_ATTR lnet_counters_t;
240 #define LNET_NI_STATUS_UP 0x15aac0de
241 #define LNET_NI_STATUS_DOWN 0xdeadface
242 #define LNET_NI_STATUS_INVALID 0x00000000
244 #define LNET_MAX_INTERFACES 16
247 * Objects maintained by the LNet are accessed through handles. Handle types
248 * have names of the form lnet_handle_xx, where xx is one of the two letter
249 * object type codes ('eq' for event queue, 'md' for memory descriptor, and
250 * 'me' for match entry). Each type of object is given a unique handle type
251 * to enhance type checking.
253 #define LNET_WIRE_HANDLE_COOKIE_NONE (-1)
255 typedef struct lnet_handle_eq {
260 * Invalidate eq handle \a h.
262 static inline void LNetInvalidateEQHandle(struct lnet_handle_eq *h)
264 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
268 * Check whether eq handle \a h is invalid.
270 * \return 1 if handle is invalid, 0 if valid.
272 static inline int LNetEQHandleIsInvalid(struct lnet_handle_eq h)
274 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
277 typedef struct lnet_handle_md {
282 * Invalidate md handle \a h.
284 static inline void LNetInvalidateMDHandle(struct lnet_handle_md *h)
286 h->cookie = LNET_WIRE_HANDLE_COOKIE_NONE;
290 * Check whether eq handle \a h is invalid.
292 * \return 1 if handle is invalid, 0 if valid.
294 static inline int LNetMDHandleIsInvalid(struct lnet_handle_md h)
296 return (LNET_WIRE_HANDLE_COOKIE_NONE == h.cookie);
299 typedef struct lnet_handle_me {
306 typedef struct lnet_process_id {
314 /** \addtogroup lnet_me
318 * Specifies whether the match entry or memory descriptor should be unlinked
319 * automatically (LNET_UNLINK) or not (LNET_RETAIN).
321 typedef enum lnet_unlink {
327 * Values of the type enum lnet_ins_pos are used to control where a new match
328 * entry is inserted. The value LNET_INS_BEFORE is used to insert the new
329 * entry before the current entry or before the head of the list. The value
330 * LNET_INS_AFTER is used to insert the new entry after the current entry
331 * or after the last item in the list.
333 typedef enum lnet_ins_pos {
334 /** insert ME before current position or head of the list */
336 /** insert ME after current position or tail of the list */
338 /** attach ME at tail of local CPU partition ME list */
344 /** \addtogroup lnet_md
348 * Defines the visible parts of a memory descriptor. Values of this type
349 * are used to initialize memory descriptors.
351 typedef struct lnet_md {
353 * Specify the memory region associated with the memory descriptor.
354 * If the options field has:
355 * - LNET_MD_KIOV bit set: The start field points to the starting
356 * address of an array of lnet_kiov_t and the length field specifies
357 * the number of entries in the array. The length can't be bigger
358 * than LNET_MAX_IOV. The lnet_kiov_t is used to describe page-based
359 * fragments that are not necessarily mapped in virtal memory.
360 * - LNET_MD_IOVEC bit set: The start field points to the starting
361 * address of an array of struct kvec and the length field specifies
362 * the number of entries in the array. The length can't be bigger
363 * than LNET_MAX_IOV. The struct kvec is used to describe fragments
364 * that have virtual addresses.
365 * - Otherwise: The memory region is contiguous. The start field
366 * specifies the starting address for the memory region and the
367 * length field specifies its length.
369 * When the memory region is fragmented, all fragments but the first
370 * one must start on page boundary, and all but the last must end on
376 * Specifies the maximum number of operations that can be performed
377 * on the memory descriptor. An operation is any action that could
378 * possibly generate an event. In the usual case, the threshold value
379 * is decremented for each operation on the MD. When the threshold
380 * drops to zero, the MD becomes inactive and does not respond to
381 * operations. A threshold value of LNET_MD_THRESH_INF indicates that
382 * there is no bound on the number of operations that may be applied
387 * Specifies the largest incoming request that the memory descriptor
388 * should respond to. When the unused portion of a MD (length -
389 * local offset) falls below this value, the MD becomes inactive and
390 * does not respond to further operations. This value is only used
391 * if the LNET_MD_MAX_SIZE option is set.
395 * Specifies the behavior of the memory descriptor. A bitwise OR
396 * of the following values can be used:
397 * - LNET_MD_OP_PUT: The LNet PUT operation is allowed on this MD.
398 * - LNET_MD_OP_GET: The LNet GET operation is allowed on this MD.
399 * - LNET_MD_MANAGE_REMOTE: The offset used in accessing the memory
400 * region is provided by the incoming request. By default, the
401 * offset is maintained locally. When maintained locally, the
402 * offset is incremented by the length of the request so that
403 * the next operation (PUT or GET) will access the next part of
404 * the memory region. Note that only one offset variable exists
405 * per memory descriptor. If both PUT and GET operations are
406 * performed on a memory descriptor, the offset is updated each time.
407 * - LNET_MD_TRUNCATE: The length provided in the incoming request can
408 * be reduced to match the memory available in the region (determined
409 * by subtracting the offset from the length of the memory region).
410 * By default, if the length in the incoming operation is greater
411 * than the amount of memory available, the operation is rejected.
412 * - LNET_MD_ACK_DISABLE: An acknowledgment should not be sent for
413 * incoming PUT operations, even if requested. By default,
414 * acknowledgments are sent for PUT operations that request an
415 * acknowledgment. Acknowledgments are never sent for GET operations.
416 * The data sent in the REPLY serves as an implicit acknowledgment.
417 * - LNET_MD_KIOV: The start and length fields specify an array of
419 * - LNET_MD_IOVEC: The start and length fields specify an array of
421 * - LNET_MD_MAX_SIZE: The max_size field is valid.
422 * - LNET_MD_BULK_HANDLE: The bulk_handle field is valid.
425 * - LNET_MD_KIOV or LNET_MD_IOVEC allows for a scatter/gather
426 * capability for memory descriptors. They can't be both set.
427 * - When LNET_MD_MAX_SIZE is set, the total length of the memory
428 * region (i.e. sum of all fragment lengths) must not be less than
431 unsigned int options;
433 * A user-specified value that is associated with the memory
434 * descriptor. The value does not need to be a pointer, but must fit
435 * in the space used by a pointer. This value is recorded in events
436 * associated with operations on this MD.
440 * A handle for the event queue used to log the operations performed on
441 * the memory region. If this argument is a NULL handle (i.e. nullified
442 * by LNetInvalidateHandle()), operations performed on this memory
443 * descriptor are not logged.
445 struct lnet_handle_eq eq_handle;
447 * The bulk MD handle which was registered to describe the buffers
448 * either to be used to transfer data to the peer or receive data
449 * from the peer. This allows LNet to properly determine the NUMA
450 * node on which the memory was allocated and use that to select the
451 * nearest local network interface. This value is only used
452 * if the LNET_MD_BULK_HANDLE option is set.
454 struct lnet_handle_md bulk_handle;
457 /* Max Transfer Unit (minimum supported everywhere).
458 * CAVEAT EMPTOR, with multinet (i.e. routers forwarding between networks)
459 * these limits are system wide and not interface-local. */
460 #define LNET_MTU_BITS 20
461 #define LNET_MTU (1 << LNET_MTU_BITS)
463 /** limit on the number of fragments in discontiguous MDs */
464 #define LNET_MAX_IOV 256
467 * Options for the MD structure. See struct lnet_md::options.
469 #define LNET_MD_OP_PUT (1 << 0)
470 /** See struct lnet_md::options. */
471 #define LNET_MD_OP_GET (1 << 1)
472 /** See struct lnet_md::options. */
473 #define LNET_MD_MANAGE_REMOTE (1 << 2)
474 /* unused (1 << 3) */
475 /** See struct lnet_md::options. */
476 #define LNET_MD_TRUNCATE (1 << 4)
477 /** See struct lnet_md::options. */
478 #define LNET_MD_ACK_DISABLE (1 << 5)
479 /** See struct lnet_md::options. */
480 #define LNET_MD_IOVEC (1 << 6)
481 /** See struct lnet_md::options. */
482 #define LNET_MD_MAX_SIZE (1 << 7)
483 /** See struct lnet_md::options. */
484 #define LNET_MD_KIOV (1 << 8)
485 /** See struct lnet_md::options. */
486 #define LNET_MD_BULK_HANDLE (1 << 9)
488 /* For compatibility with Cray Portals */
489 #define LNET_MD_PHYS 0
491 /** Infinite threshold on MD operations. See struct lnet_md::threshold */
492 #define LNET_MD_THRESH_INF (-1)
495 * A page-based fragment of a MD.
498 /** Pointer to the page where the fragment resides */
499 struct page *kiov_page;
500 /** Length in bytes of the fragment */
501 unsigned int kiov_len;
503 * Starting offset of the fragment within the page. Note that the
504 * end of the fragment must not pass the end of the page; i.e.,
505 * kiov_len + kiov_offset <= PAGE_SIZE.
507 unsigned int kiov_offset;
511 /** \addtogroup lnet_eq
515 * Six types of events can be logged in an event queue.
517 typedef enum lnet_event_kind {
518 /** An incoming GET operation has completed on the MD. */
521 * An incoming PUT operation has completed on the MD. The
522 * underlying layers will not alter the memory (on behalf of this
523 * operation) once this event has been logged.
527 * A REPLY operation has completed. This event is logged after the
528 * data (if any) from the REPLY has been written into the MD.
531 /** An acknowledgment has been received. */
534 * An outgoing send (PUT or GET) operation has completed. This event
535 * is logged after the entire buffer has been sent and it is safe for
536 * the caller to reuse the buffer.
539 * - The LNET_EVENT_SEND doesn't guarantee message delivery. It can
540 * happen even when the message has not yet been put out on wire.
541 * - It's unsafe to assume that in an outgoing GET operation
542 * the LNET_EVENT_SEND event would happen before the
543 * LNET_EVENT_REPLY event. The same holds for LNET_EVENT_SEND and
544 * LNET_EVENT_ACK events in an outgoing PUT operation.
548 * A MD has been unlinked. Note that LNetMDUnlink() does not
549 * necessarily trigger an LNET_EVENT_UNLINK event.
555 #define LNET_SEQ_GT(a, b) (((signed long)((a) - (b))) > 0)
558 * Information about an event on a MD.
560 typedef struct lnet_event {
561 /** The identifier (nid, pid) of the target. */
562 struct lnet_process_id target;
563 /** The identifier (nid, pid) of the initiator. */
564 struct lnet_process_id initiator;
565 /** The source NID on the initiator. */
566 struct lnet_process_id source;
568 * The NID of the immediate sender. If the request has been forwarded
569 * by routers, this is the NID of the last hop; otherwise it's the
570 * same as the source.
573 /** Indicates the type of the event. */
574 enum lnet_event_kind type;
575 /** The portal table index specified in the request */
576 unsigned int pt_index;
577 /** A copy of the match bits specified in the request. */
579 /** The length (in bytes) specified in the request. */
580 unsigned int rlength;
582 * The length (in bytes) of the data that was manipulated by the
583 * operation. For truncated operations, the manipulated length will be
584 * the number of bytes specified by the MD (possibly with an offset,
585 * see struct lnet_md). For all other operations, the manipulated length
586 * will be the length of the requested operation, i.e. rlength.
588 unsigned int mlength;
590 * The handle to the MD associated with the event. The handle may be
591 * invalid if the MD has been unlinked.
593 struct lnet_handle_md md_handle;
595 * A snapshot of the state of the MD immediately after the event has
596 * been processed. In particular, the threshold field in md will
597 * reflect the value of the threshold after the operation occurred.
601 * 64 bits of out-of-band user data. Only valid for LNET_EVENT_PUT.
606 * Indicates the completion status of the operation. It's 0 for
607 * successful operations, otherwise it's an error code.
611 * Indicates whether the MD has been unlinked. Note that:
612 * - An event with unlinked set is the last event on the MD.
613 * - This field is also set for an explicit LNET_EVENT_UNLINK event.
618 * The displacement (in bytes) into the memory region that the
619 * operation used. The offset can be determined by the operation for
620 * a remote managed MD or by the local MD.
621 * \see struct lnet_md::options
625 * The sequence number for this event. Sequence numbers are unique
628 volatile unsigned long sequence;
632 * Event queue handler function type.
634 * The EQ handler runs for each event that is deposited into the EQ. The
635 * handler is supplied with a pointer to the event that triggered the
636 * handler invocation.
638 * The handler must not block, must be reentrant, and must not call any LNet
639 * API functions. It should return as quickly as possible.
641 typedef void (*lnet_eq_handler_t)(struct lnet_event *event);
642 #define LNET_EQ_HANDLER_NONE NULL
645 /** \addtogroup lnet_data
649 * Specify whether an acknowledgment should be sent by target when the PUT
650 * operation completes (i.e., when the data has been written to a MD of the
653 * \see struct lnet_md::options for the discussion on LNET_MD_ACK_DISABLE
654 * by which acknowledgments can be disabled for a MD.
656 typedef enum lnet_ack_req {
657 /** Request an acknowledgment */
659 /** Request that no acknowledgment should be generated. */